Reproduce “TexSpot: 3D Texture Enhancement with Spatially-uniform Point Latent Representation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ziteng Lu, Yushuang Wu, Chongjie Ye, Yuda Qiu, Jing Shao, Xiaoyang Guo, Jiaqing Zhou, Tianlei Hu and 2 more
TexSpot enhances coarse or corrupted 3D mesh textures through a spatially uniform Texlet representation that combines locally unwrapped image patches with their 3D positions and surface normals. Its VAE-plus-diffusion pipeline improves both texture fidelity and perceptual consistency, reaching 30.04 PSNR, 0.8386 SSIM, 0.0324 LPIPS, and 44.02 FID on the reported benchmark, outperforming the compared methods.
Key insights
- The Texlet representation provides a middle ground between 2D UV textures and dense point textures: local patches retain image detail while their 3D organization supports surface-wide consistency.
- Pretrained 2D diffusion VAE features improve reconstruction and generalization compared with a conventional convolutional patch encoder, indicating that image priors remain useful when embedded in a 3D-aware pipeline.
- Increasing the representation to 8,192 patches improves VAE reconstruction to 34.21 PSNR and 0.9320 SSIM, while larger patch counts yield limited additional benefit.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.